This method is interesting in that it provides a new application for subspace clustering. However the subspace clustering method used (GPCA) has poor performance compared to newer methods like Sparse/Low-Rank Subspace Clustering. I'd like to see it done with one of these more modern approaches.
I wonder if any government agencies are providing grants for this project. Obviously this is me being a paranoiac (which I'm not normally) but this seems like an excellent start to identifying individuals whose speech you don't like.